报告时间:2018年1月24日(周三)下午14:00
报告地点:浙江大学紫金港校区蒙民伟楼402会议室
报告题目:移动平台上的物体识别
报告人:朱梦龙博士
个人简介:
Dr. Menglong Zhu is currently a Senior Software Engineer in Google Research. He obtained his PhD from University of Pennsylvania, under the supervision of Dr. Kostas Daniliidis. He graduated from Fudan with a Bachelor's degree. His research interests include object recognition, 3D pose estimation, deep learning and robotics.
报告摘要:
Mobile phone has become the largest computational platform in the world. This talk presents several key technologies that enable efficient on-device realtime deep learning applications. Mobilenet and its improvement will be introduced as backbones for various mobile applications. We will show how to achieve 10x speedup versus YOLOv2 with a lightweight detection network design we termed SSDLite. The detection framework is further extended into video domain with convolutional LSTM for feature propergation across frames. This talk will also address how to train quantized model with high accuracy. Some infrastructure and tools for mobile vision application will also be discussed.